**Description**
add auto selecting algorithm support for cudnn functions.
**Major Revision**
- add auto selecting algorithm support for cudnn functions in source
code
- add 'auto_algo' option in benchmark
- add related test
**Description**
Model benchmarks can stop due to `num_steps` or `duration` config which
will take effect when the value is set greater than 0.
If both are set greater than 0, the earliest condition reached will
work.
**Description**
Cherry-pick bug fixes from v0.8.0 to main.
**Major Revisions**
* Monitor - Fix the cgroup version checking logic (#502)
* Benchmark - Fix matrix size overflow issue in cuBLASLt GEMM (#503)
* Fix wrong torch usage in communication wrapper for Distributed
Inference Benchmark (#505)
* Analyzer: Fix bug in python3.8 due to pandas api change (#504)
* Bug - Fix bug to get metric from cmd when error happens (#506)
* Monitor - Collect realtime GPU power when benchmarking (#507)
* Add num_workers argument in model benchmark (#511)
* Remove unreachable condition when write host list (#512)
* Update cuda11.8 image to cuda12.1 based on nvcr23.03 (#513)
* Doc - Fix wrong unit of cpu-memory-bw-latency in doc (#515)
* Docs - Upgrade version and release note (#508)
Co-authored-by: guoshzhao <guzhao@microsoft.com>
Co-authored-by: Ziyue Yang <ziyyang@microsoft.com>
Co-authored-by: Yuting Jiang <yutingjiang@microsoft.com>
**Description**
This PR adds a micro-benchmark of distributed model inference workloads.
**Major Revision**
- Add a new micro-benchmark dist-inference.
- Add corresponding example and unit tests.
- Update configuration files to include this new micro-benchmark.
- Update micro-benchmark README.
---------
Co-authored-by: Peng Cheng <chengpeng5555@outlook.com>
**Description**
Cherry-pick bug fixes from v0.7.0 to main.
**Major Revisions**
* Benchmarks - Fix missing include in FP8 benchmark (#460)
* Fix bug in TE BERT model (#461)
* Doc - Update benchmark doc (#465)
* Bug: Fix bug for incorrect datatype judgement in cublas-function
source code (#464)
* Support `sb deploy` without pulling image (#466)
* Docs - Upgrade version and release note (#467)
Co-authored-by: Russell J. Hewett <russell.j.hewett@gmail.com>
Co-authored-by: Yuting Jiang <yutingjiang@microsoft.com>
**Major Revision**
- Add an option for pattern to generate mpi_pattern.txt file if
specified the path.
- In mpi pattern, serial_index and parallel_index will add in each
benchmark as environment variables.
**Minor Revision**
- Fix typo
Support FP8 in PyTorch BERT models:
* add fp8 hybrid/e4m3/e5m2 in precision arguments
* build BERT encoders with `te.TransformerLayer` to repalce
`transformers.BertModel`
* wrap forward steps with fp8 autocast
**Description**
Add correctness check in cublas-function benchmark.
**Major Revision**
- add python code of correctness check in cublas-function benchmark and test
**Description**
Add stdout logging util module and enable real-time logging flushing in executor
**Major Revision**
- Add stdout logging util module to redirect stdout into file log
- enable stdout logging in executor to write benchmark output into both stdout and file `sb-bench.log`
- enable real-time log flushing in run_command of microbenchmarks through config `log_flushing`
**Minor Revision**
- add log_n_step args to enable regular step time log in model benchmarks
- udpate related docs
Add non-zero return code for `sb deploy` and `sb run` command when
there're Ansible failures in control plane.
Return code is set to count of failure.
For failures caused by benchmarks, return code is still set per benchmark
in results json file.
Update version to include revision hash and date in "{last tag}+g{git
hash}.d{date}" format, here're the examples:
* exact tag: 0.6.0
* commit after tag: 0.6.0+gcbb1b34
* commit after tag with local changes: 0.6.0+gcbb1b34.d20221028
**Description**
Cherry-pick bug fixes from v0.6.0 to main.
**Major Revisions**
* Enable latency test in ib traffic validation distributed benchmark (#396)
* Enhance parameter parsing to allow spaces in value (#397)
* Update apt packages in dockerfile (#398)
* Upgrade colorlog for NO_COLOR support (#404)
* Analyzer - Update error handling to support exit code of sb result diagnosis (#403)
* Analyzer - Make baseline file optional in data diagnosis and fix bugs (#399)
* Enhance timeout cleanup to avoid possible hanging (#405)
* Auto generate ibstat file by pssh (#402)
* Analyzer - Format int type and unify empty value to N/A in diagnosis output file (#406)
* Docs - Upgrade version and release note (#407)
* Docs - Fix issues in document (#408)
Co-authored-by: Yang Wang <yangwang1@microsoft.com>
Co-authored-by: Yuting Jiang <yutingjiang@microsoft.com>
**Description**
Add support to store values of metrics in data diagnosis.
Take the following rules as example:
```
nccl_store_rule:
categories: NCCL_DIS
store: True
metrics:
- nccl-bw:allreduce-run0/allreduce_1073741824_busbw
- nccl-bw:allreduce-run1/allreduce_1073741824_busbw
- nccl-bw:allreduce-run2/allreduce_1073741824_busbw
- nccl-bw:allreduce-run3/allreduce_1073741824_busbw
- nccl-bw:allreduce-run4/allreduce_1073741824_busbw
nccl_rule:
function: multi_rules
criteria: 'lambda label:True if min(label["nccl_store_rule"].values())/max(label["nccl_store_rule"].values())<0.95 else False'
categories: NCCL_DIS
```
**nccl_store_rule** will store the values of the metrics in dict and save them into `label["nccl_store_rule"]` , and then **rccl_rule** can use the values of metrics through `label["nccl_store_rule"].values()` in criteria
**Description**
Add support for both jsonl and json format in data diagnosis.
**Major Revision**
- Add support for both jsonl and json format in data diagnosis
**Minor Revision**
- change related doc
- add jsonl support in cli
**Description**
Rename field in data diagnosis to be more readable.
**Major Revision**
- rename fields according to diagnosis/metric format
**Minor Revision**
- change type of diagnosis/issue_num to be int
**Description**
Add failure check feature in data diagnosis.
**Major Revision**
- Add failure check rule op to support that if there exists metric_regex not been matched by any metric in result, label as failedtest
- Split performance issue and failedtest in categories
**Minor Revision**
- replace DataFrame.append() with pd.concat since append() will be removed in later version of pandas
* Support topo-aware IB performance validation
Add a new pattern `topo-aware`, so the user can run IB performance
test based on VM's topology information. This way, the user can
validate the IB performance across VM pairs with different distance
as a quick test instead of pair-wise test.
To run with topo-aware pattern, user needs to specify three required
(and two optional) parameters in YAML config file:
--pattern topo-aware
--ibstat path to ibstat output
--ibnetdiscover path to ibnetdiscover output
--min_dist minimum distance of VM pairs (optional, default 2)
--max_dist maximum distance of VM pairs (optional, default 6)
The newly added topo_aware module then parses the topology
information, builds a graph, and generates the VM pairs with
the specified distance (# hops).
The specified IB test will then be running across these
generated VM pairs.
Signed-off-by: Jie Zhang <jessezhang1010@gmail.com>
* Add description about topology aware ib traffic tests
Signed-off-by: Jie Zhang <jessezhang1010@gmail.com>
* Add unit test to verify generated topology aware config file
This commit adds unit test to verify the generated topology aware
config file is correct. To do so, four new data files are added in
order to invoke gen_topo_aware_config function to generate topology
aware config file, then compares it with the expected config file.
Signed-off-by: Jie Zhang <jessezhang1010@gmail.com>
* Fix lint issue on Azure pipeline
Signed-off-by: Jie Zhang <jessezhang1010@gmail.com>
Fix an unexpected result value (`-0.125`) issue in ib traffic benchmark when encountering `-1` in raw output
* Check if the value is valid before the base conversion
* Add a test case to cover this situation
Fix potential port conflict due to race condition between time-to-check
to time-to-use, by binding the port all through.
Modify the function to resolve flake8 C901 while keeping the logic same.
Fix several issues in ib validation benchmark:
* continue running when timeout in the middle, instead of aborting whole mpi process
* make timeout parameter configurable, set default to 120 seconds
* avoid mixture of stdio and iostream when print to stdout
* set default message size to 8M which will saturate ib in most cases
* fix hostfile path issue so that it can be auto found in different cases
Support `node_num: 1` in mpi mode, so that we can run mpi benchmarks in
both 1 node and all nodes in one config by changing `node_num`.
Update docs and add test case accordingly.
Fix several issues in ib loopback benchmark:
* use `--report_gbits` and divide by 8 to get GB/s, previous results are
MiB/s / 1000
* use the ib_write_bw binary built in third_party instead of system path
* update the metrics name so that different hca indices have same metric
**Description**
Support multiple IB/GPU devices run simultaneously in ib validation benchmark.
**Major Revisions**
- Revise ib_validation_performance.cc so that multiple processes per node could be used to launch multiple perftest commands simultaneously. For each node pair in the config, number of processes per node will run in parallel.
- Revise ib_validation_performance.py to correct file paths and adjust parameters to specify different NICs/GPUs/NUMA nodes.
- Fix env issues in Dockerfile for end-to-end test.
- Update ib-traffic configuration examples in config files.
- Update unit tests and docs accordingly.
Closes#326.
**Description**
Support `sb run` on host directly without Docker
**Major Revisions**
- Add `--no-docker` argument for `sb run`.
- Run on host directly if `--no-docker` if specified.
- Update docs and tests correspondingly.
**Description**
Fix bugs in data diagnosis.
**Major Revision**
- add support to get baseline of the metric which uses custom benchmark naming with ':' like 'nccl-bw:default/allreduce_8_bw:0'
- save raw data of all metrics rather than metrics defined in diagnosis_rules.yaml when output_all is True
- fix bug of using wrong column index when applying format(red color and percentile) in the excel
**Description**
Cherry-pick bug fixes from v0.5.0 to main.
**Major Revisions**
* Bug - Force to fix ort version as '1.10.0' (#343)
* Bug - Support no matching rules and unify the output name in result_summary (#345)
* Analyzer - Support regex in annotations of benchmark naming for metrics in rules (#344)
* Bug - Fix bugs in sync results on root rank for e2e model benchmarks (#342)
* Bug - Fix bug of duration feature for model benchmarks in distributed mode (#347)
* Docs - Upgrade version and release note (#348)
Co-authored-by: Yuting Jiang <v-yutjiang@microsoft.com>
**Description**
Integrage result summary and update output format of data diagnosis.
**Major Revision**
- integrage result summary
- add md and html format for data diagnosis
**Description**
Use config `log_raw_data` to control whether log the raw data into file or not. The default value is `no`. We can set it as `yes` for some particular benchmarks to save the raw data into file, such as NCCL/RCCL test.
Yuting Jiang
Yifan Xiong
guoshzhao
Ziyue Yang
rafsalas19
Yang Wang
Jie Zhang
Yuting Jiang